ELM数据处理和调参经历 调参方法

After I build the ELM model(class), I should find the best result for my data.
The most important thing to do is to process my data and find suitable parameters for my model.

Butterworth Filter

I use Butterworth Filter to filter my noise torque from the robot.
One important parameter is cutoff frequency.
too big -> overfitting
too small -> too smooth

Because only torque of joint 1 is always bigger than 40, my idea is that cutoff freq of joint 1 can be bigger than other joints.
I will get it in range cf = [1, 10].
Application of the Butterworth Filter can read my artical : Butterworth Filter 巴特沃斯滤波器的python代码应用

Some function of tuning the parameters from internet

• How to tune the regularization parameter in extreme learning machine?

if you want to get better training accuracy in basic SLFN trained by ELM you should do this:
1- increase the number of neurons one by one using incremental ELM until the networks give a stable accuracy
2- after that you can change the formula of beta “B=pinv(H)*T” to this one
(B=(inv((I/C)+H’*H)*H’*T);
where I is the identity matrix you can type "help eye " in Matlab;
and try yo make your hidden layers as a squared matrix.
and after that, you can increase your regularization parameter C incrementally until you get the minimum training error.

That means, firstly you should search L, and then search lambda.

• 8 Proven Ways for improving the “Accuracy” of a Machine Learning Model

  Actually, I read it from here (How to improve the classification accuracy for machine learning).
  There are 8 ways to imporve:

  1. Add more data
  2. Treat missing and Outlier values
  3. Feature Engineering
  4. Feature Selection
  5. Multiple algorithms
  6. Algorithm Tuning
  7. Ensemble methods
  8. Cross Validation

• 基于遗传算法和正则化极限学习机的 PM2.5浓度预测研究

  I can learn the process of the experiment from here, especially in “3.3.2. 模型参数的设置”

• 融合空谱特征和集成超限学习机的高光谱图像分类